README.md

---
license: unknown
language:
- en
pretty_name: Delhi Grid Load & Weather (Apr-Aug 2024)
tags:
- time-series
- energy
- electricity
- power-grid
- load-forecasting
- weather
- india
- delhi
task_categories:
- time-series-forecasting
- tabular-regression
size_categories:
- 1K<n<10K
configs:
- config_name: power
  data_files: Delhi.csv
- config_name: weather
  data_files: Weather_Delhi.csv
---

# Delhi Grid Load & Weather (Apr–Aug 2024)

Two time-aligned CSV files covering the Delhi electricity grid's operational
state and local weather, sampled every few minutes over a 115-day span in
2024 (55 of which have data — see **Coverage and gaps** below):

- **`Delhi.csv`** — grid/power-system readings: instantaneous load,
  scheduled load, drawal, over/under-drawal, in-state generation, and grid
  frequency, plus same-day and previous-day operational summaries.
- **`Weather_Delhi.csv`** — co-located weather observations: temperature,
  humidity, wind, cloud cover, and a categorical weather description.

Both files share a `TIME STAMP` column and are meant to be inner-joined on
it. This dataset backs a physics-informed multi-horizon load/frequency
forecasting project; the raw files here are exactly as collected, with no
resampling, imputation, or feature engineering applied.

## Dataset structure

| File | Rows | Columns | Time range |
|---|---|---|---|
| `Delhi.csv` | 7,751 | 17 | 2024-04-22 10:07:28 → 2024-08-14 19:55:42 |
| `Weather_Delhi.csv` | 7,750 | 14 | 2024-04-22 10:07:28 → 2024-08-14 19:55:42 |

Joining on `TIME STAMP` (inner join) yields 7,750 aligned rows (1 timestamp
in `Delhi.csv` has no weather match).

### `Delhi.csv` fields

| Column | Type | Description |
|---|---|---|
| `TIME STAMP` | datetime string | Join key; `YYYY-MM-DD HH:MM:SS` |
| `currentfrequency` | float | Grid frequency, Hz (observed range 49.59–50.38; nominal 50 Hz) |
| `dsm_rate` | int | Demand Side Management rate/regime code; only 2 values observed in this file (`0`, `401`) |
| `load` | int | Total instantaneous grid load, MW (observed range 3,725–8,636) |
| `scheduled_load` | int | Scheduled/contracted load, MW |
| `load_drawal` | int | Power actually drawn from the grid, MW |
| `od_ud` | int | Over/under-drawal, MW — signed (`load_drawal - scheduled_load`); can be negative |
| `generation_load` | int | Local/in-state generation component, MW |
| `max_load_today`, `min_load_today` | int | Running max/min `load` for the current day as of this reading (operational, as-of statistic — not a fixed daily value) |
| `max_load_today_time`, `min_load_today_time` | string | Time-of-day (`HH:MM:SS`) those extrema occurred |
| `max_load_yesterday`, `min_load_yesterday` | int | Previous day's max/min `load` |
| `max_load_yesterday_time`, `min_load_yesterday_time` | string | Time-of-day those extrema occurred |
| `filled_at` | string | `HH:MM` ingestion/logging marker; not a physical measurement |

`load_drawal = scheduled_load + od_ud` holds exactly in this file (verified
to a residual of 0.0 across all 7,751 rows). `load ≈ load_drawal +
generation_load` holds approximately (mean residual 0.04 MW, std 4.56 MW,
97% of rows within ±10 MW, max absolute deviation 174 MW).

### `Weather_Delhi.csv` fields

| Column | Type | Description |
|---|---|---|
| `TIME STAMP` | datetime string | Join key, same format as `Delhi.csv` |
| `weather_description` | string | Categorical condition (17 unique values observed: `clear sky`, `haze`, `few clouds`, `scattered clouds`, `broken clouds`, `overcast clouds`, `mist`, `dust`, `drizzle`, `light intensity drizzle`, `light rain`, `moderate rain`, `heavy intensity rain`, `very heavy rain`, `thunderstorm`, `thunderstorm with light rain`, `thunderstorm with rain`) |
| `weather_temp` | float | Temperature, °C (observed range 24.96–45.05; inferred from plausible range for Delhi Apr–Aug, not explicitly labeled in source) |
| `weather_feels_like` | float | Apparent temperature, °C (observed range 25.96–49.96) |
| `weather_temp_min`, `weather_temp_max` | float | Local min/max temperature at observation time; near-duplicate of `weather_temp` at most timestamps |
| `weather_temp_pressure` | int | Atmospheric pressure, hPa |
| `weather_temp_humidity` | int | Relative humidity, % (0–100) |
| `weather_temp_visibility` | int | Visibility, meters |
| `weather_temp_sunrise`, `weather_temp_sunset` | int | Sunrise/sunset time, Unix epoch seconds |
| `weather_wind_speed` | float | Wind speed, m/s (observed range 0–6.69) |
| `weather_wind_deg` | int | Wind direction, degrees |
| `weather_clouds_all` | int | Cloud cover, % (0–100) |

Column naming and value ranges are consistent with the OpenWeatherMap
Current Weather API schema; this is an inference from the data's shape, not
a confirmed attribution — verify before relying on it.

## Coverage and gaps

- Native sampling is irregular: median interval 11.1 minutes, with a
  10th–90th percentile band of 1.3–11.2 minutes (i.e., frequent
  sub-minute bursts mixed with the ~11-minute baseline).
- Only 55 of the 115 calendar days in the nominal date range actually have
  data. The largest single gap is approximately 862 hours (~36 days).
  **Do not treat this as a continuous time series** — segment first on
  any gap larger than your tolerance before windowing or interpolating.
- `-9999` is used elsewhere in this data family as a missing-value sentinel
  (in both numeric and string form), but **no `-9999` values are present in
  either file as currently exported** — still worth checking for
  defensively in any downstream pipeline, since the exporter that produced
  these files may emit it under different conditions.
- `max_load_today` / `min_load_today` (and their `*_yesterday` counterparts)
  are as-of operational summaries computed by the source system at read
  time, not fixed daily aggregates — don't use them as a leakage-free
  daily max/min without checking what portion of the day had elapsed at
  each timestamp.

## Usage

Load either config with the `datasets` library (each is a single `train`
split, since the source is one CSV per config):

```python
from datasets import load_dataset

power = load_dataset("happyman11/Delhi-SLDC", "power", split="train")
weather = load_dataset("happyman11/Delhi-SLDC", "weather", split="train")

print(power[0])
print(power.features)
```

To reproduce the inner join on `TIME STAMP` described above, purely with
`Dataset.map`/`Dataset.filter` (no pandas):

```python
weather_by_time = {row["TIME STAMP"]: row for row in weather}
weather_cols = [c for c in weather.column_names if c != "TIME STAMP"]

def attach_weather(example):
    match = weather_by_time.get(example["TIME STAMP"])
    extra = {c: match[c] for c in weather_cols} if match else {c: None for c in weather_cols}
    return {**example, **extra, "_matched": match is not None}

# every row needs the same schema for Dataset.map -- unmatched rows get None
# in the weather columns rather than omitting the keys, or map() raises a
# schema-mismatch error partway through the batch that happens to contain
# the one power-only timestamp with no weather match.
joined = power.map(attach_weather).filter(lambda ex: ex["_matched"]).remove_columns("_matched")
print(len(joined))  # 7,750
```

## Dataset creation

**Source (inferred, not independently confirmed):** the power/grid file's
column set (`currentfrequency`, `load`, `scheduled_load`, `load_drawal`,
`od_ud`, `generation_load`, `dsm_rate`) matches the real-time data published
by the Delhi State Load Despatch Centre (SLDC); the weather file's schema
matches a standard current-weather API response for Delhi. Neither source
is confirmed by metadata in the files themselves — treat this section as
a best-effort inference, not a citation.

**Collection process:** unknown beyond what's inferable from the data
(apparent periodic polling of the two sources, joined only by shared
timestamp, no documented collection code included in this repository).

## Considerations for using this data

- **License / redistribution rights are not established.** This dataset
  card is shipped with `license: unknown` deliberately. If the power data
  originates from Delhi SLDC and the weather data from a commercial weather
  API, both sources likely have their own terms of use governing
  redistribution. **Confirm you have the right to redistribute this data
  before making this repository public or using it beyond personal /
  research purposes.**
- No personally identifiable information is present — this is
  aggregate grid telemetry and weather data.
- Only 55 observed days, all within April–August 2024: this supports
  short-term, same-season forecasting research on the observed period only,
  not claims about seasonal, annual, or year-over-year patterns.

## Licensing information

Not specified. See "Considerations for using this data" above.

## Citation

No canonical citation is available for this raw export. If you use this
dataset, please describe its provenance (Delhi SLDC + weather API, as
inferred above) and link back to wherever you obtained it.